It’s the Zwift draft effect which is huge. On solo efforts it’s more realistic.

With the draft A mop rider can easily average 25mph in Zwift even if they can’t even break 29 in real world solo.

And people say the average speed on Zwift is unrealistic. How often do you get to ride completely flat roads, with perfect tarmac and zero wind. When you do you'll average similar to Zwift.

For solo riding, this is largely true

However, this still does not account for the insanely low CdA conferred by riding in a group of 10+, where the first few riders generate comparable power. IRL, a solo rider doing 330 W (at 70 kg) will pull away from a group where the leader pulls at 260 W (at 70 kg). On Zwift, if a group is driven by five riders doing 260 W (at 70 kg), that group will reel back the solo rider (I've been there on both sides of things).

However, this still does not account for the insanely low CdA conferred by riding in a group of 10+, where the first few riders generate comparable power. IRL, a solo rider doing 330 W (at 70 kg) will pull away from a group where the leader pulls at 260 W (at 70 kg). On Zwift, if a group is driven by five riders doing 260 W (at 70 kg), that group will reel back the solo rider (I've been there on both sides of things).

I don’t think the wattage is close.
For example I finished a 2 hr ride with an avg of 216 watts, in the charts others around and top of me we’re only 3 or 4 above 200 other around 160, one even 127! Have to say I only have the Numbers from my vector2s.
Seems weird to me but as long as I get a ride in I couldn’t care less.

Have you noticed a difference in the races that have full draft mode? Presumably the riders 'sitting in' need to generate less power so the average for the group is lower but it's not clear if the speed of the group is determined by the average power of the group or the lead rider.

Test Parameters
All of the test riders were set to 183cm height, 75kg weight, and rode Zwift Carbon bikes with 32mm Zwift wheels.
Interesting side-note: Zwift’s draft effect actually takes rider height and weight into account – similar to outdoors! So a taller rider will create a stronger draft than a shorter rider, and a heavier rider a stronger draft than a light one. Zwift computes an estimate of a rider’s frontal area and uses this to compute the wind resistance they encounter, as well as the draft “wake” they produce.
Test Methodology
Tests were done in “Meetup-Only View” on Watopia’s Tempus Fugit route because it’s the flattest on Zwift, and it has a timed section (Fuego Flats Reverse, 4.4 miles long) which could be used to measure the speeds of each test formation.
All of the tests were done with four riders. Because I ran out of ANT sticks and computers!


Tests and Results
Test 1

All riders @ 300W
Segment time 10:14.8
Speed: 41.34 kph (25.64 mph)

Notes:

All four riders continually “churned” on the front, alternating between poking their nose into the wind, then getting slowed so another rider could come around to the front. This is what you typically see at the front of a non-TT Zwift race.

-----------------------------------------------------------------------------------------------------------------------------------
Test 3

Next, I tried to guess what wattage the lead rider would need to hold in order to beat the time set by the churning group of 300W monsters from test 1. I settled on 320 watts – here are the results:

Rider 1 @ 320W, Rider 2 @244W, Rider 3 @ 225W, Rider 4 @ 209W
Segment time: 10:22.4
Speed: 39.83 kph (25.33 mph)

Notes:

Riders received power savings of 24%, 30%, and 35% (2nd, 3rd, and 4th rider respectively). This lines up with the power savings seen in other tests.
In a TTT situation with all riders taking equal pulls on the front at these wattages, each rider would average 250W.
This group didn’t quite beat the 300W churn, but it was close.

----------------------------------------------------------------------------------------------------------------------------------------

Test 4

This time around I wanted to make sure we beat the 300W churn group. So I bumped the lead rider up to 350W:

Rider 1 @ 350W, Rider 2 @270W, Rider 3 @ 244W, Rider 4 @ 235W
Segment time: 10:01.2
Speed: 42.28 kph (26.23 mph)

Notes:

Riders received power savings of 23%, 30%, and 33% (2nd, 3rd, and 4th rider respectively). This lines up with the power savings seen in other tests.
In a TTT situation with all riders taking equal pulls on the front at these wattages, each rider would average 275W. This is crucial to understand: that even with Zwift’s “speed churning” from test 1, the four riders in this test significantly beat test 1’s time by riding efficiently in single file formation.

I'm quicker in reality than on zwift.
Probably because I'm smaller and have lower drag than what Zwift predicts.

I have the same impression.

How does the multi body simulation of Zwift work? How is the calculation work distributed between server an clients? Anyone here who knows the simulation techniques of such video games?

I suppose that the clients do solve the „local“ equation of motion. In order to get the interaction with others, the clients get every time step a list of riders in the neighborhood. Depending of the number of neighbors the client can calculate something like the lokal air density, i.e. in a group everyone cycles in a „cloud“ of low air density. But inside this cloud every one his individual cda, mass, location, speed, gradient etc.